The pnv_get_proc_freq() function was recently changed to call
cpufreq_get(), instead of cpufreq_quick_get(), in order to fetch
a more up-to-date value for the CPU frequency:
cd77b5ce208c153260ed7882d8910f2395bfaabd
powerpc/powernv/cpufreq: Fix the frequency read by /proc/cpuinfo
Unfortunately, this function is called from show_cpuinfo() in
arch/powerpc/kernel/setup-common.c with preemption disabled. The
cpufreq_get() function might do a down_read(), which can sleep.
With CONFIG_DEBUG_KERNEL and CONFIG_DEBUG_ATOMIC_SLEEP set, a warning
like the following is generated when running 'cat /proc/cpuinfo':
BUG: sleeping function called from invalid context at
kernel/locking/rwsem.c:23
in_atomic(): 1, irqs_disabled(): 0, pid: 16939, name: cat
CPU: 33 PID: 16939 Comm: cat Tainted: G W 4.15.0-smp-DEV
#1
Call Trace:
[c000000fef07bab0] [c000000000a32c30] dump_stack+0xb0/0xf0 (unreliable)
[c000000fef07baf0] [c0000000001343a8] ___might_sleep+0x178/0x1b0
[c000000fef07bb70] [c000000000a50f58] down_read+0x38/0x90
[c000000fef07bba0] [c0000000008287d0] cpufreq_get+0x50/0xc0
[c000000fef07bbf0] [c000000000097a08] pnv_get_proc_freq+0x28/0x60
[c000000fef07bc20] [c00000000002c554] show_cpuinfo+0x194/0x450
[c000000fef07bcb0] [c00000000039c848] seq_read+0x1f8/0x590
[c000000fef07bd40] [c00000000040d1d4] proc_reg_read+0xb4/0x180
[c000000fef07bd90] [c00000000035d6a0] vfs_read+0x100/0x220
[c000000fef07bde0] [c00000000035dc6c] SyS_read+0x6c/0x110
[c000000fef07be30] [c00000000000b220] system_call+0x58/0x6c
On Mon, 2018-01-08 at 21:30 -0800, John Sperbeck wrote:
The pnv_get_proc_freq() function was recently changed to call
cpufreq_get(), instead of cpufreq_quick_get(), in order to fetch
a more up-to-date value for the CPU frequency:
cd77b5ce208c153260ed7882d8910f2395bfaabd
powerpc/powernv/cpufreq: Fix the frequency read by /proc/cpuinfo
Unfortunately, this function is called from show_cpuinfo() in
arch/powerpc/kernel/setup-common.c with preemption disabled. The
cpufreq_get() function might do a down_read(), which can sleep.
With CONFIG_DEBUG_KERNEL and CONFIG_DEBUG_ATOMIC_SLEEP set, a warning
like the following is generated when running 'cat /proc/cpuinfo':
We could just either remove the preempt_disable completely like
x86 and keep it racy, or stick a cpus_read_lock around it. I dont think
we need that preempt_disable, it's definitely overkill.
Michael, what do you think ? I'm keen on sync'ing with x86 here...
Cheers,
Ben.
BUG: sleeping function called from invalid context at kernel/locking/rwsem.c:23
in_atomic(): 1, irqs_disabled(): 0, pid: 16939, name: cat
CPU: 33 PID: 16939 Comm: cat Tainted: G W 4.15.0-smp-DEV #1
Call Trace:
[c000000fef07bab0] [c000000000a32c30] dump_stack+0xb0/0xf0 (unreliable)
[c000000fef07baf0] [c0000000001343a8] ___might_sleep+0x178/0x1b0
[c000000fef07bb70] [c000000000a50f58] down_read+0x38/0x90
[c000000fef07bba0] [c0000000008287d0] cpufreq_get+0x50/0xc0
[c000000fef07bbf0] [c000000000097a08] pnv_get_proc_freq+0x28/0x60
[c000000fef07bc20] [c00000000002c554] show_cpuinfo+0x194/0x450
[c000000fef07bcb0] [c00000000039c848] seq_read+0x1f8/0x590
[c000000fef07bd40] [c00000000040d1d4] proc_reg_read+0xb4/0x180
[c000000fef07bd90] [c00000000035d6a0] vfs_read+0x100/0x220
[c000000fef07bde0] [c00000000035dc6c] SyS_read+0x6c/0x110
[c000000fef07be30] [c00000000000b220] system_call+0x58/0x6c
Benjamin Herrenschmidt [off-list ref] writes:
On Mon, 2018-01-08 at 21:30 -0800, John Sperbeck wrote:
quoted
The pnv_get_proc_freq() function was recently changed to call
cpufreq_get(), instead of cpufreq_quick_get(), in order to fetch
a more up-to-date value for the CPU frequency:
cd77b5ce208c153260ed7882d8910f2395bfaabd
powerpc/powernv/cpufreq: Fix the frequency read by /proc/cpuinfo
Unfortunately, this function is called from show_cpuinfo() in
arch/powerpc/kernel/setup-common.c with preemption disabled. The
cpufreq_get() function might do a down_read(), which can sleep.
With CONFIG_DEBUG_KERNEL and CONFIG_DEBUG_ATOMIC_SLEEP set, a warning
like the following is generated when running 'cat /proc/cpuinfo':
This was reported by Nick back in November just a few days after the
patch went in, so it's a little disappointing that it's still broken.
We could just either remove the preempt_disable completely like
x86 and keep it racy, or stick a cpus_read_lock around it. I dont think
we need that preempt_disable, it's definitely overkill.
Michael, what do you think ? I'm keen on sync'ing with x86 here...
Yeah I think we can drop it. The cpufreq seems to have some locking that
looks like it'll probably work - famous last words.
Who's writing the patch?
cheers